Neural Audio Coding with Linear Filters for Low-Bitrate Speech
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing linear predictive-based codecs struggle to meet the requirements of delivering high-quality speech coding at low bit-rates with low complexity and low memory, while neural network-based codecs are resource-intensive and computationally expensive.
Innovation Solution
Combining a non-linear source generated by a first machine learning system with a linear filter estimated by a second machine learning system, such as neural networks, to perform audio coding, reducing complexity and enhancing quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If neural network-based codecs are used, then audio quality is improved, but computational complexity and resource requirements increase
Solution Approach 1:
The audio coding system is segmented into two independent neural network components: one dedicated to generating the excitation signal and another dedicated to generating the filter parameters. This segmentation allows each component to be optimized separately and processed independently, reducing the overall computational burden while maintaining high audio quality through specialized processing in each segment.
Solution Approach 2:
The invention extracts and separates the traditionally coupled processes of excitation signal generation and filter parameter generation into distinct neural network modules. By taking out these functions and processing them separately, the system achieves better computational efficiency and resource utilization while preserving the high-quality audio reconstruction capability.
2Device complexity
If linear predictive-based codecs are used, then computational complexity is reduced, but audio quality deteriorates
Solution Approach 1:
The invention replaces traditional mechanical linear predictive coding algorithms with neural network-based models for generating excitation signals and filter parameters. This substitution introduces non-linear processing capabilities that significantly improve audio quality while maintaining computational efficiency through the inherent parallel processing capabilities of neural networks.
Solution Approach 2:
The system combines the strengths of traditional linear predictive coding (efficient filter structure) with neural network capabilities (non-linear signal processing) to create a hybrid audio coding system. This composite approach leverages the computational efficiency of linear models while incorporating the quality-enhancing non-linear processing of neural networks.
3Manufacturing precision
If high bit-rates are used, then audio quality is improved, but data transmission requirements increase
Solution Approach 1:
The invention changes the parameter representation by using neural networks to generate compact excitation signals and filter parameters that capture the essential characteristics of the audio signal. This parameter transformation enables high-quality audio reconstruction at lower bit-rates by efficiently encoding the most important signal features.
Solution Approach 2:
The neural network-based excitation signal generator creates a compact representation that copies the essential characteristics of the original audio signal's excitation component. This efficient copying mechanism allows for high-fidelity audio reconstruction using significantly fewer bits compared to traditional methods.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Systems and techniques are described for coding audio signals. For example, a voice decoder can generate, using a first neural network, an excitation signal for at least one sample of an audio signal at least in part by performing a non-linear operation based on one or more inputs to the first neural network, the excitation signal being configured to excite a learned linear filter. The voice decoder can further generate, using the learned linear filter and the excitation signal, at least one sample of a reconstructed audio signal. For example, a second neural network can be used to generate coefficients for one or more learned linear filters, which receive as input the excitation signal generated by the first neural network trained to perform the non-linear operation.